ChatGPT’s reported citation mix is shifting toward sources that originate evidence: government health bodies, research sources, regulators and hands-on testing sites. That does not establish a universal rule for every model, prompt or AI answer, but it is a meaningful signal for GEO teams: being a recognised publisher is not the same thing as being the source an answer engine needs to cite.
The practical implication is not “stop publishing explainers.” It is that an explainer-led visibility strategy is exposed when it merely restates material available from an upstream agency, vendor, dataset or researcher. The stronger GEO asset is increasingly the page containing evidence, method, product truth or observed experience that another publisher cannot reproduce.
What the citation data actually says
Search Engine Journal’s report on Ahrefs data says PubMed rose from 64th to fourth among domains most frequently cited in U.S. ChatGPT answers between March and September. In the same comparison, WebMD fell from 18th to 79th. Government health, research and regulator sites made up much of the group that climbed, while health and finance explainers lost ground. searchenginejournal.com
The direction is more important than any one domain. Ahrefs characterises WebMD and Verywell Health as sites that summarise research, and groups Healthline and Investopedia with health and finance explainers. Its reported gainers also included regulators such as the CFPB and FTC, as well as Google and Microsoft support pages. searchenginejournal.com
Ahrefs’ own interpretation is that ChatGPT “seems to be prioritizing original sources over sites that summarize them.” searchenginejournal.com That is a useful working hypothesis, not a settled law of AI search. The underlying comparison covered the 200 domains most frequently cited in U.S. answers, based on 2.4 million answers in March and 3.8 million in September. searchenginejournal.com
There is an important methodological caveat. Ahrefs adds prompts to its tracking system over time, so rank changes could reflect a changing question set rather than a change in citation selection. To address that, it compared roughly 25,000 prompts that appeared in both months among its 150,000 highest-volume prompts, and reported that every site in its chart moved in the same direction on those shared prompts. searchenginejournal.com That strengthens the directional finding, but it still does not make this a complete view of every query class or every assistant.
Why publisher authority is no longer enough
Traditional publishing authority can still matter, but this data suggests it is not a sufficient GEO proposition when the page is a polished middle layer. If a model can cite the regulator that issued the notice, the research source that produced the finding, or the vendor whose documentation defines a feature, a secondary summary has to offer something more than readability.
That “something more” should be concrete. The useful distinction is not primary source versus secondary source as a purity test. It is replaceable versus non-replaceable. A page becomes more difficult to replace when it contains material that exists because your organisation did the work, operates the product, measured the environment or maintains the record.
Ken Ashe’s analysis of the reporting makes the point plainly: explainers retain value for human readers, particularly where primary sources are dense or difficult to apply, but they need to add judgment rather than simply compress upstream information. It identifies original measurements, firsthand testing, methodology, product documentation, changelogs and real usage comparisons as examples of material that can be cite-worthy. kenashe.ai
For GEO, this reframes the content brief. The question should not begin with “What comprehensive article can we write around this topic?” Start with: “What fact, test, operational detail or maintained reference can only this site publish?” The article can follow. The unique artifact is the asset that gives the article a reason to be selected.
Build evidence assets, not just answer-shaped pages
A team does not need to become a research institution to respond to this signal. But it should stop treating all informational production as interchangeable. A high-volume explainer programme and a primary-evidence programme may both create pages, yet they make very different claims on an answer engine’s citation choices.
- Original datasets: publish measurements your organisation collected, with enough context to understand what was counted and how.
- Methodology: state how a comparison, benchmark or test was performed rather than presenting a conclusion without its basis.
- Firsthand testing: document real product use, observed results and the limits of the test.
- Product documentation: maintain accurate pages about capabilities, pricing, availability and implementation when you are the vendor or operator.
- Operational records: use dated field notes, changelogs, maintained source lists and other records tied to work your organisation actually performs.
These are not guaranteed citation mechanisms. The supplied evidence does not show that adding a methodology section, publishing a dataset or revising documentation causes a ChatGPT citation. It does, however, support a more defensible allocation of GEO effort than producing another summary of an existing source.
The right standard is also higher than making a page sound authoritative. As Ashe argues, AI citation is not won by authority signals alone; the durable advantage is being the page that authoritative sources cannot replace. kenashe.ai For a GEO practitioner, that is a content-model decision before it is an optimisation decision.
Do not overread the search behaviour data
The citation trend is notable partly because the reported web-search rate is low in the sampled environment. Out of 66.6 million free, logged-out responses in Ahrefs’ August-to-September fan-out sample, 14.4% involved a web search. The data covered all countries and a single model snapshot, according to the report. searchenginejournal.com
That limits the conclusions we can draw. Citation-rank data is not traffic data, and a source that is more often cited in answers is not necessarily receiving more visits. Search Engine Journal explicitly notes that Ahrefs’ citation ranks do not measure traffic. searchenginejournal.com A GEO dashboard that treats mentions, citations, referral sessions and commercial outcomes as one metric will hide the difference.
The report also says 6.8% of 15 million fan-out queries were site: searches, spread over 116,000 domains. Among site: searches aimed at roughly 160 of the most-searched domains, 75.9% went to brand sites as defined by the report. searchenginejournal.com That makes owned documentation especially relevant for brands: where the system seeks a specific site, the authoritative product or company source has an obvious role.
What to change in a GEO review
Review important informational pages against the sources ChatGPT currently cites for the corresponding topics. If the cited result is a regulator, research database, vendor document or tester, identify whether your page contributes original value beyond that source. If not, do not assume stronger prose, more subheadings or more generic coverage will close the gap.
Prioritise pages where the business can add a real primary artifact. That could mean turning internal product knowledge into maintained documentation, publishing a repeatable test method, exposing a useful data table, or recording observations from actual work. Then monitor whether your pages are cited, what claims they are cited for and which alternative sources persist.
Finally, keep the conclusion proportionate. This is a reported pattern from Ahrefs data, not proof that explainer content has ceased to work. Search Engine Journal notes that the pattern could change with future model updates and recommends checking which sources ChatGPT cites for core topics, especially when OpenAI updates its model. searchenginejournal.com The durable lesson is narrower and more useful: GEO strategy should invest in evidence that begins on your site, not only interpretation that arrives after everyone else’s.